Problem
Enterprise knowledge lives in a dozen systems, each with its own permission model, freshness, and format. Point-solution RAG against one of them works — until the second, third, and fourth source need the same treatment, each with different authorization rules that a single flat search index can't represent.
Architecture
SharePoint · Confluence · GitHub · Jira · Backstage · Enterprise DBs · APIs · Documents
Sources remain systems of record — the fabric never becomes a second copy of truth.
Ingestion / Integration
Connector per source, incremental sync (change-data-capture or webhooks) where the source supports it, rather than full re-crawl by default.
Metadata / Classification
Every item gets a sensitivity tier, an owner, and a freshness timestamp before it's indexed — the single design decision everything downstream depends on.
Knowledge Processing
Chunking, entity extraction, and structure normalization tuned per source type, not a single generic pipeline for every format.
Hybrid Retrieval
Lexical + vector (+ graph where relationships matter) — pure vector search under-performs on exact-match enterprise queries like ticket IDs or policy numbers.
Security-Aware Retrieval
Document- and field-level authorization enforced at query time, using the requesting user's actual entitlements from the enterprise identity provider — not a static "public index" filtered after the fact.
LLM / Agent Layer
Grounded generation with enforced citation back to the specific retrieved passage — never asserted from model memory alone.
Enterprise Applications
Copilots, agents, search, and workflow automation — all consuming the same governed fabric rather than each building its own index.
AQEVON architectural principles
- Source systems remain systems of record — the fabric indexes and governs access, it does not replace them.
- Metadata and ownership are established before indexing, not after.
- Enterprise identity is reused — no parallel permission system.
- Authorization is enforced at retrieval time, per query.
- Hybrid retrieval is preferred over vector-only retrieval for enterprise workloads.
- Generated answers are grounded and cited back to source content.
- Evaluation is continuous — a standing labeled query set, not a one-time launch test.
- Observability separates query, retrieval, and authorization signals into distinct log streams.
Components
Source connectors, a metadata/classification store, a hybrid (lexical + vector) index, an authorization-aware retrieval gateway sitting in front of the LLM/agent layer, and a citation-enforcement step on generation.
Data flow
Source of truth never leaves the origin system. The fabric ingests, classifies, and indexes a governed copy for retrieval — access revocation at the source must propagate to the fabric, not just block future ingestion.
Security
The most common failure mode this architecture is designed against: a system that authenticates the user correctly but retrieves from an index with no document-level authorization, surfacing content the user shouldn't see. Authorization is a retrieval-layer concern here, not an application-layer afterthought.
Trade-offs
Freshness vs. index cost (more frequent re-indexing costs more but reduces staleness); a centralized fabric vs. per-team indexes (centralization governs better but requires more upfront cross-team coordination).
Cost considerations
For many enterprise knowledge workloads, ingestion and re-indexing can become a larger cost driver than query volume — a common budgeting mistake is sizing this like a query-cost problem when it's actually, often, a pipeline-cost problem.
Evaluation
Retrieval precision/recall against a labeled query set, plus a dedicated authorization-leak test — verifying a user genuinely cannot retrieve content outside their entitlements, not just that the UI doesn't display it.
Production considerations
Incremental re-indexing at scale, source-system rate limits, and access-revocation propagation latency (how long after a permission is revoked at the source does the fabric actually stop serving that content?) are the three considerations most often underestimated.
Possible implementation
This is presented as an AQEVON Reference Architecture — a pattern to adapt to a specific environment, not a claimed production customer implementation. Actual implementation details (which vector store, which identity provider integration) are environment-specific decisions made during an architecture engagement.
Lessons learned Illustrative Scenario
The most common risk with this pattern is treating it as a search project instead of an authorization project — teams that get retrieval quality right first and authorization "later" tend to ship a fast, fluent system that leaks data it shouldn't, discovered only once someone asks the wrong question and gets the right-sounding wrong answer.
AQEVON's point of view
RAG is a retrieval technique. Knowledge architecture is a discipline. Most enterprises need the second one and are currently only building the first — this reference architecture is what closes that gap.